Contrastive independent component analysis Scientists constantly face this challenge: you have data from two conditions—patients vs. healthy controls, treated vs. untreated cells, one species vs. another—and you need to find patterns unique to one group. Standard approaches either analyze datasets separately or lump them together, missing the comparative signal entirely. Contrastive PCA addressed this by subtracting a scaled version of the background covariance from the foreground, then finding principal components of the difference. The problem: researchers must manually choose how much background to subtract. Different choices yield different patterns, requiring parameter sweeps and subjective judgment about which result looks best. Probabilistic contrastive PCA improved the statistical framework but inherited the same tuning burden. Kexin Wang, Aida Maraj, and Anna Seigal propose contrastive independent component analysis (cICA), which eliminates this guesswork entirely. The key insight: by analyzing fourth-order statistics rather than second-order covariances, the method can determine exactly how much each background pattern contributes to the foreground—no manual tuning required. Moreover, each background component can contribute differently, rather than assuming a single global scaling factor as previous methods do. The algorithm uses a hierarchical tensor decomposition on fourth-order cumulants, recursively extracting eigenstructure to recover the mixing patterns. This higher-order approach also inherits a fundamental advantage from classical independent component analysis: the patterns become uniquely identifiable, whereas PCA-based methods can only recover them up to rotations. Results confirm the theoretical advantages translate to practice. On synthetic data, cICA recovers true foreground patterns with over 0.9 cosine similarity, substantially outperforming cPCA and probabilistic cPCA at their best hyperparameter settings. On mouse protein data distinguishing Down syndrome, cICA achieves the highest clustering score. Applied to human versus primate gene expression, the algorithm's top patterns prioritize genes independently validated as human-specific evolutionary drivers. The practical upshot: when your question is "what's different about this condition?", moving beyond covariance to higher-order structure reveals patterns that simpler methods miss—and does so automatically. Paper: pnas.org/doi/10.1073/pnas.24…